Evaluating and Controlling Agents

1. Evaluation Grows a Dimension

Everything from the eval lesson still holds here: golden sets, judges, human anchors, online truth. Agents just add one more dimension on top. The unit of quality is no longer an output, but a trajectory: a sequence of decisions, tool calls, and recoveries that ends, or doesn't, in a completed task. Two systems can both technically "resolve the ticket." One took four clean steps. The other took nineteen, called the refund tool three separate times, and just got lucky